kris/nobs
1
0
Fork 0
mirror of https://github.com/henrydcase/nobs.git synced 2024-11-22 15:18:57 +00:00
Code Issues Releases Wiki Activity

makes AddReduced503 25% faster

Browse Source
This commit is contained in:
Henry Case 2018-11-20 12:34:51 +00:00
parent ea2ffa2d61
commit e621ca22b3
1 changed files with 404 additions and 449 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

93
dh/sidh/p503/arith_amd64.s
View File

@ -139,78 +139,33 @@ TEXT ·fp503AddReduced(SB),NOSPLIT,$0-24
MOVQ x+8(FP), REG_P1 MOVQ x+8(FP), REG_P1
MOVQ y+16(FP), REG_P2 MOVQ y+16(FP), REG_P2
// Used later to calculate a mask
XORQ CX, CX
// [R8-R15]: z = x + y // [R8-R15]: z = x + y
MOVQ ( 0)(REG_P1), R8 MOVQ ( 0)(REG_P1), R8; ADDQ ( 0)(REG_P2), R8; MOVQ R8, ( 0)(REG_P3)
MOVQ ( 8)(REG_P1), R9 MOVQ ( 8)(REG_P1), R9; ADCQ ( 8)(REG_P2), R9; MOVQ R9, ( 8)(REG_P3)
MOVQ (16)(REG_P1), R10 MOVQ (16)(REG_P1), R10; ADCQ (16)(REG_P2), R10; MOVQ R10, (16)(REG_P3)
MOVQ (24)(REG_P1), R11 MOVQ (24)(REG_P1), R11; ADCQ (24)(REG_P2), R11; MOVQ R11, (24)(REG_P3)
MOVQ (32)(REG_P1), R12 MOVQ (32)(REG_P1), R12; ADCQ (32)(REG_P2), R12; MOVQ R12, (32)(REG_P3)
MOVQ (40)(REG_P1), R13 MOVQ (40)(REG_P1), R13; ADCQ (40)(REG_P2), R13; MOVQ R13, (40)(REG_P3)
MOVQ (48)(REG_P1), R14 MOVQ (48)(REG_P1), R14; ADCQ (48)(REG_P2), R14; MOVQ R14, (48)(REG_P3)
MOVQ (56)(REG_P1), R15 MOVQ (56)(REG_P1), R15; ADCQ (56)(REG_P2), R15; MOVQ R15, (56)(REG_P3)
ADDQ ( 0)(REG_P2), R8
ADCQ ( 8)(REG_P2), R9
ADCQ (16)(REG_P2), R10
ADCQ (24)(REG_P2), R11
ADCQ (32)(REG_P2), R12
ADCQ (40)(REG_P2), R13
ADCQ (48)(REG_P2), R14
ADCQ (56)(REG_P2), R15
MOVQ P503X2_0, AX MOVQ ·p503x2+ 0(SB), AX; SUBQ AX, R8
SUBQ AX, R8 MOVQ ·p503x2+ 8(SB), AX; SBBQ AX, R9
MOVQ P503X2_1, AX MOVQ ·p503x2+16(SB), AX; SBBQ AX, R10
SBBQ AX, R9 MOVQ ·p503x2+24(SB), AX; SBBQ AX, R11
SBBQ AX, R10 MOVQ ·p503x2+32(SB), AX; SBBQ AX, R12
MOVQ P503X2_3, AX MOVQ ·p503x2+40(SB), AX; SBBQ AX, R13
SBBQ AX, R11 MOVQ ·p503x2+48(SB), AX; SBBQ AX, R14
MOVQ P503X2_4, AX MOVQ ·p503x2+56(SB), AX; SBBQ AX, R15
SBBQ AX, R12
MOVQ P503X2_5, AX
SBBQ AX, R13
MOVQ P503X2_6, AX
SBBQ AX, R14
MOVQ P503X2_7, AX
SBBQ AX, R15
SBBQ $0, CX // mask MOVQ ( 0)(REG_P3), AX; CMOVQCC R8, AX; MOVQ AX, ( 0)(REG_P3)
MOVQ ( 8)(REG_P3), AX; CMOVQCC R9, AX; MOVQ AX, ( 8)(REG_P3)
// move z to REG_P3 MOVQ (16)(REG_P3), AX; CMOVQCC R10, AX; MOVQ AX, (16)(REG_P3)
MOVQ R8, ( 0)(REG_P3) MOVQ (24)(REG_P3), AX; CMOVQCC R11, AX; MOVQ AX, (24)(REG_P3)
MOVQ R9, ( 8)(REG_P3) MOVQ (32)(REG_P3), AX; CMOVQCC R12, AX; MOVQ AX, (32)(REG_P3)
MOVQ R10, (16)(REG_P3) MOVQ (40)(REG_P3), AX; CMOVQCC R13, AX; MOVQ AX, (40)(REG_P3)
MOVQ R11, (24)(REG_P3) MOVQ (48)(REG_P3), AX; CMOVQCC R14, AX; MOVQ AX, (48)(REG_P3)
MOVQ R12, (32)(REG_P3) MOVQ (56)(REG_P3), AX; CMOVQCC R15, AX; MOVQ AX, (56)(REG_P3)
MOVQ R13, (40)(REG_P3)
MOVQ R14, (48)(REG_P3)
MOVQ R15, (56)(REG_P3)
// if z<0 add p503x2 back
MOVQ P503X2_0, R8
MOVQ P503X2_1, R9
MOVQ P503X2_3, R10
MOVQ P503X2_4, R11
MOVQ P503X2_5, R12
MOVQ P503X2_6, R13
MOVQ P503X2_7, R14
ANDQ CX, R8
ANDQ CX, R9
ANDQ CX, R10
ANDQ CX, R11
ANDQ CX, R12
ANDQ CX, R13
ANDQ CX, R14
ADDQ R8, ( 0)(REG_P3)
ADCQ R9, ( 8)(REG_P3)
ADCQ R9, (16)(REG_P3)
ADCQ R10,(24)(REG_P3)
ADCQ R11,(32)(REG_P3)
ADCQ R12,(40)(REG_P3)
ADCQ R13,(48)(REG_P3)
ADCQ R14,(56)(REG_P3)
RET RET